AI is software built by people, not a living creature. AI includes many methods for prediction, recognition, search, and generation. This lesson focuses on autoregressive language models: they generate text one token at a time.
Technical writing calls some plausible but incorrect or unsupported outputs hallucinations. This term describes an output problem; it does not mean the program has a mind or experiences human hallucinations. We can avoid personifying a machine while still identifying and correcting its output errors.
A reasoning mode may spend more computation on intermediate steps, search, or checking before producing an answer. This can help on some tasks but does not guarantee correctness. A product label such as thinking is not evidence of a human-like conscious experience.
We will build a small word-bigram model. It counts what follows one word and turns those counts into probabilities. This makes conditional prediction visible, but it is a teaching model, not the full algorithm used by modern neural language models.
A neural language model learns numerical weights from training examples. At generation time, those weights and the available context determine scores for candidate tokens. The model does not simply look up a stored internet sentence or count only the word immediately before it.
Imagine a made-up distribution: blue 70%, gray 20%, orange 10%. These numbers sum to 100%; they are an illustration, not measurements from an AI service. Sampling draws according to the probabilities. Greedy selection always picks a highest-probability candidate in this fixed example.
Sampling can produce different answers, and it can repeat an answer. A low positive temperature concentrates the distribution; it does not make facts truer. Even temperature zero does not promise identical output across real systems: context, model versions, tools, and numerical execution can also change results.
A fluent or high-probability answer is not necessarily true. A model can produce an unsupported claim in a confident style. Ask what evidence supports the claim, rather than treating fluent wording as proof.
Probability describes a choice. Evidence supports a factual claim. Keep these two jobs separate.
Treat AI as a tool, without assuming human feelings or intentions. People remain responsible for designing and using systems appropriately. Identifying a model's limits helps us choose a better tool or checking method; it is not a moral judgment about a machine.
Wrong outputs can come from model limitations, training data, missing context, stale retrieved material, or software defects. A clear question can help, but cannot guarantee a correct answer. Check the cause before blaming the learner or assuming one more prompt will solve everything.
Training changes weights; using a model is inference. Supplying a document in a prompt changes the context, not necessarily the weights. Retrieval can bring useful sources into that context, but those sources and the generated claims still need checking.
Pick one claim from an answer. Open its original source, check the date and context, and compare what the source actually says. Recalculate numbers or run an appropriate test when possible. Ask for a hint, then explain the checked result in your own words.
Your first experiment: after one word, candidate A appears three times and candidate B once. What are the probabilities? Predict them before opening the builder, then compare greedy selection with repeated sampling.
Build a small model, observe its limits, and check what you learned. Understanding the mechanism is one step toward using AI well.